Scaling Inference Lab
ARIA's £50m programme for deploying novel compute and interconnect architectures at scale in UK data centres.
Last refreshed: 15 June 2026 · Appears in 1 active topic
Can photonic interconnects finally break the UK data-centre power deadlock?
Timeline for Scaling Inference Lab
Oriole moves data with light, not copper
UK Startups and InnovationBackground
The Scaling Inference Lab emerged in early 2026 when ARIA, the UK's Advanced Research and Invention Agency, committed £50m to CommonAI CIC to build a national testbed for next-generation AI compute. The lab focuses on the inference phase of AI, where most data-centre energy is consumed, and runs in six-month cycles that prove new hardware architectures under real-world workloads. Its first cluster deployed in mid-2026, partnering with Oriole Networks and AMD to run the world's first large-scale photonic AI network, reporting an 81% cut in core interconnect power .
The programme is hosted by CommonAI, a community-interest company that brings together industry, academia and government around open benchmarks. ARIA's programme director Suraj Bramhavar leads the initiative under ARIA's Scaling Compute opportunity space. A second cluster is scheduled for August 2026, focusing on heterogeneous orchestration for agentic workloads with hardware partner Callosum. Additional UK Government funding of £20m+ was announced to expand the lab's national capability.
The lab matters beyond its own budget. UK data-centre capacity in the London basin is constrained by grid limits, and every efficiency gain compounds: an operator who can run more AI compute under the same power envelope does not need to wait for new grid connections. By providing a public testbed where startups such as Oriole can prove performance claims before commercial deployment, the Scaling Inference Lab shortens the PATH from lab result to industry-wide adoption.